[llvm] [X86] Prefer `lock or` over mfence (PR #106555)

[Valentin Churavy via llvm-commits]

================
@@ -882,7 +886,8 @@ def ProcessorFeatures {
   list<SubtargetFeature> NHMTuning = [TuningMacroFusion,
                                       TuningSlowDivide64,
                                       TuningInsertVZEROUPPER,
-                                      TuningNoDomainDelayMov];
+                                      TuningNoDomainDelayMov,
+                                      TuningAvoidMFENCE];
----------------
vchuravy wrote:

I believe that's what GCC did.

```
+DEF_TUNE (X86_TUNE_AVOID_MFENCE, "avoid_mfence",
 	 m_CORE_ALL | m_BDVER | m_ZNVER | m_GENERIC)
 ```

https://github.com/llvm/llvm-project/pull/106555